🤖 AI

Modeling Vehicle-Type-Specific Pedestrian Crash Avoidance Behavior in Safety-Critical Interactions Using Smooth-Mamba Deep Reinforcement Learning

This study introduces the SMamba-DDPG framework, a Smooth-Mamba Deep Reinforcement Learning model that effectively captures vehicle-type-specific pedestrian crash avoidance behaviors, revealing that pedestrians react more quickly and cautiously to automated vehicles than to human-driven ones, thereby providing critical insights for safer automated driving system design and realistic mixed-traffic simulations.

Original authors: Qingwen Pu, Kun Xie, Hong Yang, Di Yang, Junqing Wang

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Qingwen Pu, Kun Xie, Hong Yang, Di Yang, Junqing Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy city intersection as a giant, high-stakes dance floor. On one side, you have human drivers (HDVs) who might wave, make eye contact, or honk to say, "I see you, go ahead!" On the other side, you have self-driving cars (AVs), which are like dancers who never speak, never make eye contact, and just follow a strict, silent script.

The big question this paper asks is: How does a pedestrian's dance step change depending on whether they are dancing with a chatty human or a silent robot?

Here is a breakdown of the study's journey, using simple analogies:

1. The Problem: The "Silent Partner" Effect

Most safety research focuses on how the car tries to avoid hitting the person. But this study flips the script. It asks: How does the person react when they see a robot car versus a human car?

The Gap: Humans rely on social cues (eye contact, hand waves) to know if a car will stop. Robots don't do this. The researchers wanted to know if this silence makes people freeze, hesitate, or move differently when a crash is about to happen.

2. The Tool: A "Super-Brain" for Traffic

To figure this out, the researchers didn't just watch videos; they built a digital "brain" using a new type of artificial intelligence called SMamba-DDPG.

The Analogy: Think of standard AI as a student who studies one math problem at a time. If the problem changes suddenly (like a car swerving), the student gets confused.
The Innovation: The new "SMamba" brain is like a student who can remember the whole story of the traffic scene, not just the last second. It also has a "smoothing filter" (like noise-canceling headphones) that helps it stay calm and make smooth decisions even when the situation gets chaotic and scary.
The Training: They fed this brain data from the Argoverse 2 dataset—a massive library of real-world driving videos from six US cities. They taught the brain two separate ways to behave: one way for interacting with human drivers, and a completely different way for interacting with robot drivers.

3. The Experiment: Replaying the Dance

Once the AI was trained, the researchers asked it to "replay" thousands of near-crash scenarios to see if it could mimic real human behavior.

The Result: The AI was incredibly accurate. It could predict exactly how a real person would speed up, slow down, or step sideways to avoid a collision.
The Discovery: The AI confirmed that people do react differently.
- Faster Reaction to Robots: Surprisingly, people reacted quicker to the self-driving cars than to human cars. It's as if the silence of the robot made people jump to action sooner, perhaps because they didn't want to wait for a signal that might never come.
- Slower Crossing: However, once they started moving, people walked slower across the street when facing a robot car. They were more cautious, taking smaller steps and moving more carefully.

4. The "What If" Scenarios (Counterfactuals)

The researchers ran a fun experiment: "What if we took the 'Robot-Scared' pedestrian and put them in front of a human car?"

The Outcome: The pedestrian became overly cautious, walking slower and braking more often than a real human would.
The Reverse: If they took the "Human-Scared" pedestrian and put them in front of a robot, they walked faster and were more aggressive.
The Lesson: This proves that our behavior is heavily influenced by who (or what) we are looking at. We can't just use one rulebook for all cars; we need different rules for robots and humans.

5. The Safety Verdict

Finally, the researchers used their AI to generate thousands of new "what-if" crash scenarios to see who is safer.

The Finding: Even though people are more cautious around robots, the conflict rate (the chance of a near-miss) was actually lower when interacting with self-driving cars compared to human cars.
Why? Because the self-driving cars were more consistent and predictable, the pedestrians felt they could trust the robot to stop, leading to fewer chaotic, dangerous moments.

Summary

This paper built a smart computer model to learn how pedestrians dance around cars. It found that we treat robot cars differently than human cars: we react faster to them but move more cautiously. By understanding these differences, we can design safer self-driving cars that know exactly how to interact with us, making our city streets a safer dance floor for everyone.

Technical Summary: Modeling Vehicle-Type-Specific Pedestrian Crash Avoidance Behavior Using Smooth-Mamba Deep Reinforcement Learning

1. Problem Statement

As automated vehicles (AVs) increasingly share roadways with human-driven vehicles (HDVs), a critical research gap exists in understanding how pedestrians respond differently to these distinct vehicle types during safety-critical interactions. Unlike HDV drivers, AVs lack non-verbal social cues (e.g., eye contact, gestures) that pedestrians rely on to negotiate right-of-way, potentially altering pedestrian trust, risk perception, and avoidance strategies.

Existing literature suffers from three primary limitations:

Vehicle-Centric Focus: Most studies prioritize AV safety systems (e.g., emergency braking, path planning) rather than modeling pedestrian behavioral changes during interactions.
Lack of Differentiation: Current frameworks often aggregate AV and HDV interactions, failing to capture how specific vehicle types influence pedestrian decision-making.
Neglect of Safety-Critical Events: Traditional models focus on routine crossing or early-stage approach behaviors, overlooking the sequential dynamics of emergency avoidance when collisions are imminent.

Furthermore, standard Deep Reinforcement Learning (DRL) architectures struggle with the "rare-event" nature of safety-critical data, exhibiting training instability during abrupt state transitions (e.g., sudden braking or sharp turns) and failing to capture long-range temporal dependencies effectively.

2. Methodology

2.1. Data Source and Processing

The study utilizes the Argoverse 2 Motion Forecasting Dataset, comprising 250,000 high-resolution urban driving scenarios.

Differentiation: The dataset distinguishes between the ego-vehicle (an AV equipped with visible LiDAR and cameras) and all other vehicles (HDVs).
Extraction: Safety-critical interactions were identified using the Curvilinear Time-to-Collision (CurvTTC) metric, which accounts for curved vehicle trajectories and non-linear pedestrian movements, unlike traditional straight-line TTC.
Dataset Split: 411 Pedestrian-AV and 2,404 Pedestrian-HDV safety-critical interactions were extracted. These were split into 80% for training and 20% for testing.

2.2. The SMamba-DDPG Framework

To model pedestrian crash avoidance, the authors propose the Smooth-Mamba Deep Deterministic Policy Gradient (SMamba-DDPG) framework. This architecture addresses the instability of standard DRL in rare-event learning by integrating structured state-space models with smoothing mechanisms.

Architecture: The framework employs an Actor-Critic structure where both networks embed SMamba blocks.
Mamba Integration: The Mamba (Selective State Space Model) component captures long-range temporal dependencies and filters noise more efficiently than LSTM or Transformer-based DRL.
Smoothing Mechanism: To address distributional imbalances and activation spikes common in emergency maneuvers, the authors introduce a Smooth-SiLU (S-SiLU) activation function and apply smoothing vectors to the gating and output projections within the Mamba block. This stabilizes gradient flow during abrupt state transitions without adding inference overhead.
Policy Learning:
- State Space ( $S_t$ ): Includes relative longitudinal/lateral distances, velocities, and relative velocities between the pedestrian and the vehicle.
- Action Space ( $A_t$ ): Continuous acceleration commands for the pedestrian in longitudinal and lateral directions.
- Reward Function: A two-level reward system was tested. Step-level rewards included distance-based, relative-velocity-based, and absolute-velocity-based formulations. Episode-level rewards included collision penalties and goal achievement bonuses. The absolute-velocity-based reward ( $R_V$ ) was found to yield the most stable and realistic behaviors.

2.3. Experimental Design

Separate Training: Two distinct models were trained independently: one for Pedestrian-AV interactions and one for Pedestrian-HDV interactions.
Baselines: The proposed model was compared against supervised learning models (NN, LSTM, Transformer) and standard/hybrid DRL models (DDPG, DQN, LSTM-DDPG, Transformer-DDPG, Mamba-DDPG).
Validation: The study utilized trajectory reconstruction, reaction time analysis, counterfactual scenario testing, and large-scale safety analysis using model-generated data.

3. Key Contributions

Quantitative Modeling of Vehicle-Type Specificity: This is among the first studies to quantitatively model and compare pedestrian crash avoidance behaviors specifically for AVs versus HDVs using a DRL framework.
SMamba-DDPG Framework: The development of a novel DRL architecture that integrates Mamba state-space models with smoothing mechanisms to stabilize learning in rare, safety-critical events and address distributional imbalances.
Comprehensive Analysis Framework: The establishment of a validation pipeline including reaction-time analysis, counterfactual experiments (swapping policies between vehicle types), and large-scale conflict-risk evaluation to quantify behavioral differences.

4. Results

4.1. Model Performance

Superior Accuracy: The SMamba-DDPG framework outperformed all baselines. The velocity-based variant ( $SMamba-DDPG_V$ $S M amba - D D P G_{V}$ ) achieved the lowest Root Mean Square Error (RMSE) across kinematic variables.
- Pedestrian-AV: Longitudinal Velocity RMSE = 0.62 m/s; ADE = 0.65 m; FDE = 1.12 m.
- Pedestrian-HDV: Longitudinal Velocity RMSE = 0.63 m/s; ADE = 0.64 m; FDE = 1.13 m.
Stability: The smoothing mechanism successfully reduced training variance and improved convergence compared to standard DDPG and other hybrid models.

4.2. Behavioral Differences

Reaction Times: Pedestrians responded significantly faster to AVs than to HDVs. The model accurately reproduced these human-like response delays, with no significant statistical difference between reconstructed and real-world data.
Counterfactual Analysis:
- When an AV-trained policy was applied to HDV trajectories, pedestrians exhibited lower crossing speeds (14.7% reduction) and smoother acceleration profiles, suggesting a more conservative strategy.
- When an HDV-trained policy was applied to AV trajectories, pedestrians exhibited higher crossing speeds (2.17 m/s vs. 1.08 m/s) and more assertive behavior.
Yielding Behavior: In large-scale simulations, pedestrians yielded more frequently to AVs (46.07%) compared to HDVs (37.15%), indicating greater caution or uncertainty regarding AV decision-making.

4.3. Safety Analysis

Large-scale analysis of model-generated data revealed that Pedestrian-AV interactions consistently yielded lower conflict rates than Pedestrian-HDV interactions across all speed conditions. At high speeds (>3.5 m/s), AV conflict rates were 10–15 percentage points lower than HDV rates.

5. Significance and Claims

The paper claims that its findings highlight the necessity of incorporating vehicle-type-specific pedestrian behavioral models for:

Safer AV Design: Understanding that pedestrians react differently (faster but more cautiously) to AVs can inform AV control strategies and interaction designs (e.g., external human-machine interfaces).
Realistic Traffic Simulation: Aggregating AV and HDV interactions into a single model compromises the accuracy of mixed-traffic simulations. Separate policies are required to accurately predict risk and yielding behaviors.
Risk Assessment: The study demonstrates that pedestrian risk perception is not static but is dynamically influenced by the perceived type of the approaching vehicle, with AVs inducing a more conservative (yielding) response despite faster reaction times.

The authors conclude that the SMamba-DDPG framework provides a robust method for capturing the sequential, abrupt dynamics of pedestrian avoidance, offering a pathway to more realistic and safe automated driving system deployment.